25 of the 26 tasks come from a freely available FreeCAD exercise series from the German-speaking region; one case (the bearing block, case 900) is our own construction and is therefore the only one shown in full. We publish no third-party drawings and no geometry – every third-party case is referenced solely via its source video and via SHA256 checksums (for case 013 the series links no public video; the reference runs via the series itself). The checksums document which exact state was measured against, without disclosing the data itself.
Selection: We included every part of the exercise series whose materials could yield a verifiable case — a dimensioned single-part drawing plus a reference model against which the dimensions can be re-measured. The corpus thus covers the range of the series, from simple plate and milled parts to rotationally symmetric parts and thread cases; assemblies are not included.
In brief: 26 tasks, mostly from a public FreeCAD exercise series. We do not show third-party drawings; every task is uniquely fixed via its source video and a checksum. Unfamiliar terms are explained in the glossary.
Each row is a case with its anchoring label, the number of dimensioned target dimensions and the three ceilings from the methodology. The column “Random hit (L/A)” gives the random-hit rate for lengths and angles – the null model against which a dimensional-accuracy value is to be placed. It is computed as follows: random values are distributed uniformly over the part's value range (angles over 0–180°), and the fraction of them that would find a match on the reference model is counted with the same tolerance as in the scoring. A high value arises above all on very small parts — there the minimum tolerance is relatively large, and dimensional accuracy alone is correspondingly of little probative force.
The underlying drawings and reference models are works of third parties. We have decided neither to show them nor to make them available for download; the sources deposited in the respective author's video description remain the only place to obtain the original data. On this website only metadata appears (checksums, dimension count, ceilings) and – in the case-by-case comparison – the 3D depictions of our own AI reconstructions.
More on why this is so and what an “anchoring case” means is in the FAQ and in the methodology.